How does online learning affect bias and variance in streaming ML systems?

Online learning reduces bias by adapting continuously but can increase variance due to sensitivity to recent data.

Online learning updates models incrementally as new data arrives. This allows rapid adaptation to changing distributions, reducing bias in dynamic environments. However, because updates rely heavily on recent samples, the model becomes sensitive to noise and short-term fluctuations, increasing variance.

To mitigate this, techniques like learning rate decay, mini-batching, and replay buffers are used. In non-stationary environments, online learning is essential but must be carefully regularized to avoid catastrophic forgetting.